Abnormal object detection of transmission lines plays a very important role in improving the safety, reliability and stability of transmission systems, but existing object detection has not been effectively designed for large scale changes, many small objects, dark light, partial occlusion of abnormal objects on the line, resulting that recognition speed is slow, it is easy to be disturbed by the environment, and false positives and false negatives are frequent. In response to the above problems, this paper adopts a two-stage deep network. The feature Pyiamid network(FPN) is used to extract multi-scale features so that the backbone network can better adapt to multi-scale changes of objects, and feature enhancement is performed through the global network to obtain clearer and representative multi-scale object features. A feature-guided region proposal generation network is proposed in the region proposal network(RPN), which can generate sparse and arbitrary-shaped anchors to generate tighter mask bounding boxes. In the detection stage, a multi-task loss function is used to improve prediction accuracy and generalization ability of the network, and to improve detection performance of abnormal objects. An ablation experiment and performance comparison on the MS COCO dataset verify the effectiveness and advancement of the proposed method. The detection accuracy of abnormal objects on the transmission line dataset reaches 77$%$, which is better than mainstream deep learning object detection methods.